Determination of organic arsenic acids by machine learning-assisted SERS on a silicon-modified coating of Fe3O4@Ag

Abstract

Surface-enhanced Raman spectroscopy (SERS) is a highly promising technique for the detection and identification of arsenic speciation. In this study, a core–shell substrate, Fe₃O₄@SiO₂@Ag nanoparticles, was prepared via a hydrothermal method. Characterization techniques such as TEM, XRD, UV-vis spectroscopy, FT-IR spectroscopy, and VSM analyses confirmed that silver nanoparticles (Ag NPs) were uniformly distributed on the surface of the Fe₃O₄ core, with the SiO₂ layer serving as a protective shell. The SERS of dimethylarsinic acid and roxarsone on this substrate revealed distinct As–C and As–O stretching vibrations in the range of 600–1000 cm⁻¹. The fabricated substrate exhibited a wide linear detection range (5.0–1000.0 µg L⁻¹). The limits of detection (LODs) for dimethylarsinic acid and roxarsone were 4.0 µg L⁻¹ and 0.6 µg L⁻¹, respectively, with enhancement factors (EFs) of 4.12 × 10⁷ and 2.54 × 10⁸, respectively. Furthermore, the substrate exhibited good stability, selectivity and resistance to interference from representative compounds. Machine learning models based on partial least squares discriminant analysis (PLS-DA), random forest (RF), K-nearest neighbors (KNN), and support vector machine (SVM) algorithms effectively identified the SERS spectra of single and mixed organic arsenic acids in natural water samples, achieving a classification accuracy of 99.1%. Theoretical calculations elucidated the synergistic effect of the electromagnetic and chemical mechanisms that contributed to the ultrahigh SERS activity. The photoinduced charge transfer and the formation of electromagnetic hotspots due to coupling between silver nanoparticles effectively promoted the Raman signal intensity. This approach demonstrates the potential of combining advanced machine learning models with SERS for the accurate and efficient analysis of mixed organic arsenic acids in real environmental water samples.